Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “efficiency.”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 271 records · Page 15
All sourcesNASA NTRSDOE OSTI
NASA NTRS1985

Gallium arsenide solar cell efficiency: Problems and potential

Under ideal conditions the GaAs solar cell should be able to operate at an AMO efficiency exceeding 27 percent, whereas to date the best measured efficiencies barely exceed 19 percent. Of more concern is the fact that there has been no improvement in the past half decade, despite the expenditure of considerable effort. State-of-the-art GaAs efficiency is analyzed in an attempt to determine the feasibility of improving on the status quo. The possible gains to be had in the planar cell. An attempt is also made to predict the efficiency levels that could be achieved with a grating geometry. Both the N-base and the P-base BaAs cells in their planar configurations have the potential to operate at AMO efficiencies between 23 and 24 percent. For the former the enabling technology is essentially in hand, while for the latter the problem of passivating the emitter surface remains to be solved. In the dot grating configuration, P-base efficiencies approaching 26 percent are possible with minor improvements in existing technology. N-base grating cell efficiencies comparable to those predicted for the P-base cell are achievable if the N surface can be sufficiently passivated.

Weizer, V. G.↗
NASA NTRS1985

High-efficiency silicon solar-cell design and practical barriers

A numerical evaluation technique is used to study the impact of practical barriers, such as heavy doping effects (Auger recombination, band gap narrowing), surface recombination, shadowing losses and minority-carrier lifetime (Tau), on a high efficiency silicon solar cell performance. Considering a high Tau of 1 ms, efficiency of a silicon solar cell of the hypothetical case is estimated to be around 29%. This is comparable with (detailed balance limit) maximum efficiency of a p-n junction solar cell of 30%. Value of Tau is varied from 1 second to 20 micro. Heavy doping effects, and realizable values of surface recombination velocities and shadowing, are then considered in succession and their influence on cell efficiency is evaluated and quantified. These practical barriers cause the cell efficiency to reduce from the maximum value of 29% to the experimentally achieved value of about 19%. Improvement in open circuit voltage V sub oc is required to achieve cell efficiency greater than 20%. Increased value of Tau reduces reverse saturation current and, hence, improves V sub oc. Control of surface recombination losses becomes critical at higher V sub oc. Substantial improvement in Tau and considerable reduction in surface recombination velocities is essential to achieve cell efficiencies greater than 20%.

Mokashi, A.↗
NASA NTRS1985

Summary of high-efficiency solar-cell research

High-efficiency solar-cell activities supporting efforts to achieve the DOE Five-Year Plan goals are summarized. Specific objectives are to identify and resolve key generic problems that limit cell efficiency to below theoretically predicted values and to design and fabricate cells having efficiences equal to or greater than 20% (AM1.5). Theoretical curves for various p-n junction cells were shown. The effects of practical barriers on cell efficiency was depicted along with the modeling parameters. Cell design parameters used in the analyses were described. The usefulness and present limitations of the existing modeling capabilities were presented. The historical evolution of the efficiencies of cells made from web and edge-defined film-fed growth (EFG) silicon ribbons were also described. The status of contemporary higher-efficiency technical capabilities and future activities to raise efficiencies were stated.

Kachare, R.↗
NASA NTRS1989

The 25 percent-efficient GaAs Cassegrainian concentrator cell

Very high-efficiency GaAs Cassegrainian solar cells have been fabricated in both the n-p and p-n configurations. The n-p configuration exhibits the highest efficiency at concentration, the best cells having an efficiency eta of 24.5 percent (100X, AM0, temperature T = 28 C). Although the cells are designed for operation at this concentration, peak efficiency is observed near 300 suns (eta = 25.1 percent). To our knowledge, this is the highest reported solar cell efficiency for space applications. The improvement in efficiency over that reported at the previous SPRAT conference is attributed primarily to lower series resistance and improved grid-line plating procedures. Using previously measured temperature coefficients, researchers estimate that the n-p GaAs cells should deliver approximately 22.5 percent efficiency at the operating conditions of 100 suns and T = 80 C. This performance exceeds the NASA program goal of 22 percent for the Cassegrainian cell. One hundred Cassegrainian cells have been sent to NASA as deliverables, sixty-eight in the n-p configuration and thirty-two in the p-n configuration.

Hamaker, H. C.↗
NASA NTRS1991

High-efficiency, radiation-resistant GaAs space cells

Although many GaAs solar cells are intended for space applicatons, few measurements of cell degradation after radiation are available, particularly for cells with efficiencies exceeding 20 percent (one-sun, AMO). Often the cell performance is optimized for the highest beginning-of-life (BOL) efficiency, despite the unknown effect of such design on end-of-life (EOL) efficiencies. The results of a study of the radiation effects on p-n GaAs cells are presented. The EOL efficiency of GaAs space cell can be increased by adjusting materials growth parameters, resulting in a demonstration of 16 percent EOL efficiency at one-sun, AMO. Reducing base doping levels to below 3 x 10(exp 17)/cu m and decreasing emitter thickness to 0.3 to 0.5 micron for p-n cells led to significant improvements in radiation hardness as measured by EOL/BOL efficiency ratios for irradiation of 10(exp -15)/sq cm electrons at 1 MeV. BOL efficiency was not affected by changes in emitter thickness but did improve with lower base doping.

Bertness, K. A.↗
NASA NTRS1988

Ku-band high efficiency GaAs MMIC power amplifiers

The development of Ku-band high efficiency GaAs MMIC power amplifiers is examined. Three amplifier modules operating over the 13 to 15 GHz frequency range are to be developed. The first MMIC is a 1 W variable power amplifier (VPA) with 35 percent efficiency. On-chip digital gain control is to be provided. The second MMIC is a medium power amplifier (MPA) with an output power goal of 1 W and 40 percent power-added efficiency. The third MMIC is a high power amplifier (HPA) with 4 W output power goal and 40 percent power-added efficiency. An output power of 0.36 W/mm with 49 percent efficiency was obtained on an ion implanted single gate MESFET at 15 GHz. On a dual gate MESFET, an output power of 0.42 W/mm with 27 percent efficiency was obtained. A mask set was designed that includes single stage, two stage, and three stage single gate amplifiers. A single stage 600 micron amplifier produced 0.4 W/mm output power with 40 percent efficiency at 14 GHz. A four stage dual gate amplifier generated 500 mW of output power with 20 dB gain at 17 GHz. A four-bit digital-to-analog converter was designed and fabricated which has an output swing of -3 V to +/- 1 V.

Tserng, H. Q.↗
NASA NTRS1991

The oxygen content of ocean bottom waters, the burial efficiency of organic carbon, and the regulation of atmospheric oxygen

Data for the burial efficiency of organic carbon with marine sediments have been compiled for 69 locations. The burial efficiency as here defined is the ratio of the quantity of organic carbon which is ultimately buried to that which reaches the sediment-water interface. As noted previously, the sedimentation rate exerts a dominant influence on the burial efficiency. The logarithm of the burial efficiency is linearly related to the logarithm of the sedimentation rate at low sedimentation rates. At high sedimentation rates the burial efficiency can exceed 50% and becomes nearly independent of the sedimentation rate. The residual of the burial efficiency after the effect of the sedimentation rate has been subtracted is a weak function of the O2 concentration in bottom waters. The scatter is sufficiently large, so that the effect of the O2 concentration in bottom waters on the burial efficiency of organic matter could be either negligible or a minor but significant part of the mechanism that controls the level of O2 in the atmosphere.

Non-NASA Center↗
NASA NTRS1949

Performance of Single-Stage Turbine of Mark 25 Torpedo Power Plant with Two Special Nozzles: Efficiency with Standard Rotor Blades - III

A Mark 25 torpedo power plant modified to operate as a single-stage turbine was investigated to determine the performance with two nozzle designs and a standard first-stage rotor having 0.40-inch blades with a 17O met-air angle. Both nozzles had smaller port cross-sectional areas than those nozzles of similar design, which were previously investigated. The performance of the two nozzles was compared on the basis of blade, rotor, and brake efficiencies as a function of blade-jet speed ratio for pressure ratios of 8, 15 (design), and 20. At pressure ratios of 15 and 20, the blade efficiency obtained with the nozzle having circular passages (K) was higher than that obtained with the nozzle having rectangular passages (J). At a pressure ratio of 8, the efficiencies obtained with the two nozzles were comparable for blade-jet speed ratios of less than 0.260. For blade-jet speed ratios exceeding this value, nozzle K yielded slightly higher efficiencies. The maximum blade efficiency of 0.569 was obtained with nozzle K at a pressure ratio of 8 and a blade-jet speed ratio of 0.295. At design speed and pressure ratio, nozzle K yielded a maximum blade efficiency of 0.534, an increase of 0.031 over that obtained with nozzle J. When the blade efficiencies of the two nozzles were compared with those of four other nozzles previously investigated, the maximum difference for the six nozzles with this rotor was 0.050. From, this comparison, no specific effect of nozzles size or shape on over-all performance was discernible.

Schum, Harold J.↗
NASA NTRS1996

Highly Efficient Amplifier for Ka-Band Communications

An amplifier developed under a Small Business Innovation Research (SBIR) contract will have applications for both satellite and terrestrial communications. This power amplifier uses an innovative series bias arrangement of active devices to achieve over 40-percent efficiency at Ka-band frequencies with an output power of 0.66 W. The amplifier is fabricated on a 2.0- by 3.8-square millimeter chip through the use of Monolithic Microwave Integrated Circuit (MMIC) technology, and it uses state-of-the-art, Pseudomorphic High-Electron-Mobility Transistor (PHEMT) devices. Although the performance of the MMIC chip depends on these high-performance devices, the real innovations here are a unique series bias scheme, which results in a high-voltage chip supply, and careful design of the on-chip planar output stage combiner. This design concept has ramifications beyond the chip itself because it opens up the possibility of operation directly from a satellite power bus (usually 28 V) without a dc-dc converter. This will dramatically increase the overall system efficiency. Conventional microwave power amplifier designs utilize many devices all connected in parallel from the bias supply. This results in a low-bias voltage, typically 5 V, and a high bias current. With this configuration, substantial I(sup 2) R losses (current squared times resistance) may arise in the system bias-distribution network. By placing the devices in a series bias configuration, the total current is reduced, leading to reduced distribution losses. Careful design of the on-chip planar output stage power combiner is also important in minimizing losses. Using these concepts, a two-stage amplifier was designed for operation at 33 GHz and fabricated in a standard MMIC foundry process with 0.20-m PHEMT devices. Using a 20-V bias supply, the amplifier achieved efficiencies of over 40 percent with an output power of 0.66 W and a 16-dB gain over a 2-GHz bandwidth centered at 33 GHz. With a 28-V bias, a power level of 1.1 W was achieved with a 12-dB gain and a 36-percent efficiency. This represents the best reported combination of power and efficiency at this frequency. In addition to delivering excellent power and gain, this Ka-band MMIC power amplifier has an efficiency that is 10 percent greater than existing designs. The unique design offers an excellent match for spacecraft applications since the amplifier supply voltage is closely matched to the typical value of spacecraft bus voltage. These amplifiers may be used alone in applications of 1 W or less, or several may be combined or used in an array to produce moderate power, Ka-band transmitters with minimal power combining and less thermal stress owing to the combination of excellent efficiency and power output. The higher voltage operation of this design may also save mass and power because the dc-dc power converter is replaced with a simpler voltage regulator.

Source record↗
NASA NTRS2001

Power and Efficiency Optimized in Traveling-Wave Tubes Over a Broad Frequency Bandwidth

A traveling-wave tube (TWT) is an electron beam device that is used to amplify electromagnetic communication waves at radio and microwave frequencies. TWT's are critical components in deep space probes, communication satellites, and high-power radar systems. Power conversion efficiency is of paramount importance for TWT's employed in deep space probes and communication satellites. A previous effort was very successful in increasing efficiency and power at a single frequency (ref. 1). Such an algorithm is sufficient for narrow bandwidth designs, but for optimal designs in applications that require high radiofrequency power over a wide bandwidth, such as high-density communications or high-resolution radar, the variation of the circuit response with respect to frequency must be considered. This work at the NASA Glenn Research Center is the first to develop techniques for optimizing TWT efficiency and output power over a broad frequency bandwidth (ref. 2). The techniques are based on simulated annealing, which has the advantage over conventional optimization techniques in that it enables the best possible solution to be obtained (ref. 3). Two new broadband simulated annealing algorithms were developed that optimize (1) minimum saturated power efficiency over a frequency bandwidth and (2) simultaneous bandwidth and minimum power efficiency over the frequency band with constant input power. The algorithms were incorporated into the NASA coupled-cavity TWT computer model (ref. 4) and used to design optimal phase velocity tapers using the 59- to 64-GHz Hughes 961HA coupled-cavity TWT as a baseline model. In comparison to the baseline design, the computational results of the first broad-band design algorithm show an improvement of 73.9 percent in minimum saturated efficiency (see the top graph). The second broadband design algorithm (see the bottom graph) improves minimum radiofrequency efficiency with constant input power drive by a factor of 2.7 at the high band edge (64 GHz) and increases simultaneous bandwidth by 500 MHz.

Wilson, Jeffrey D.↗
NASA NTRS2005

High-Efficiency Multi-Junction Space Solar Development Utilizing Lattice Grading

Progress towards achieving a high one-sun air mass 0 (AM0) efficiency in a monolithic dual junction solar cell comprised of a 1.62 eV InGaP top cell and a 1.1 eV InGaAs bottom cell grown on buffered GaAs is reported. The performance of stand-alone 1.62 eV InGaP and 1.1 eV InGaAs cells is compared to that of the dual junction cell. Projected AM0 efficiencies of 15.7% and 16.5% are expected for the 1.62 eV InGaP and 1.1 eV InGaAs cells grown on buffered GaAs. The dual junction cell has a projected one-sun AM0 conversion efficiency of 17%. The projected efficiencies are based upon the application of an optimized anti-reflective coating (ARC) to the as-grown cells. Quantum efficiency (QE) data obtained from the dual junction cell indicate that is is bottom cell current limited with the top cell generating 50% more current than the bottom cell. A comparison of the QE data for the stand-alone 1.1 eV InGaAs cell to that of the 1.1 eV InGaAs bottom cell in the tandem configuration indicates a degradation of the bottom cell conversion efficiency in the tandem configuration. The origin of this performance degradation is at present unknown. If the present limitation can be overcome, then a one-sun AM0 efficiency of 26% is achievable with the 1.62 eV/1.1 eV dual junction cell grown lattice-mismatched to GaAs.

Stan, Mark A.↗
NASA NTRS2013

High-Efficiency Nested Hall Thrusters for Robotic Solar System Exploration

This work describes the scaling and design attributes of Nested Hall Thrusters (NHT) with extremely large operational envelopes, including a wide range of throttleability in power and specific impulse at high efficiency (>50%). NHTs have the potential to provide the game changing performance, powerprocessing capabilities, and cost effectiveness required to enable missions that cannot otherwise be accomplished. NHTs were first identified in the electric propulsion community as a path to 100- kW class thrusters for human missions. This study aimed to identify the performance capabilities NHTs can provide for NASA robotic and human missions, with an emphasis on 10-kW class thrusters well-suited for robotic exploration. A key outcome of this work has been the identification of NHTs as nearly constant-efficiency devices over large power throttling ratios, especially in direct-drive power systems. NHT systems sized for robotic solar system exploration are predicted to be capable of high-efficiency operation over nearly their entire power throttling range. A traditional Annular Hall Thruster (AHT) consists of a single annular discharge chamber where the propellant is ionized and accelerated. In an NHT, multiple annular channels are concentrically stacked. The channels can be operated in unison or individually depending on the available power or required performance. When throttling an AHT, performance must be sacrificed since a single channel cannot satisfy the diverse design attributes needed to maintain high thrust efficiency. NHTs can satisfy these requirements by varying which channels are operated and thereby offer significant benefits in terms of thruster performance, especially under deep power throttling conditions where the efficiency of an AHT suffers since a single channel can only operate efficiently (>50%) over a narrow power throttling ratio (3:1). Designs for 10-kW class NHTs were developed and compared with AHT systems. Power processing systems were considered using either traditional Power Processing Units (PPU) or Direct Drive Units (DDU). In a PPU-based system, power from the solar arrays is transformed from the low voltage of the arrays to the high voltage needed by the thruster. In a DDU-based system, power from the solar arrays is fed to the thruster without conversion. DDU-based systems are attractive for their simplicity since they eliminate the most complex and expensive part of the propulsion system. The results point to the strong potential of NHTs operating with either PPUs or DDUs to benefit robotic and human missions through their unprecedented power and specific impulse throttling capabilities. NHTs coupled to traditional PPUs are predicted to offer high-efficiency (>50%) power throttling ratios 320% greater than present capabilities, while NHTs with direct-drive power systems (DDU) could exceed existing capabilities by 340%. Because the NHT-DDU approach is implicitly low-cost, NHT-DDU technology has the potential to radically reduce the cost of SEP-enabled NASA missions while simultaneously enabling unprecedented performance capability.

Hofer, Richard R.↗
NASA NTRS2014

Simulation of Fluid Flow and Collection Efficiency for an SEA Multi-element Probe

Numerical simulations of fluid flow and collection efficiency for a Science Engineering Associates (SEA) multi-element probe are presented. Simulation of the flow field was produced using the Glenn-HT Navier-Stokes solver. Three-dimensional unsteady results were produced and then time averaged for the heat transfer and collection efficiency results. Three grid densities were investigated to enable an assessment of grid dependence. Simulations were completed for free stream velocities ranging from 85-135 meters per second, and free stream total pressure of 44.8 and 93.1 kilopascals (6.5 and 13.5 pounds per square inch absolute). In addition, the effect of angle of attack and yaw were investigated by including 5 degree deviations from straight for one of the flow conditions. All but one of the cases simulated a probe in isolation (i.e. in a very large domain without any support strut). One case is included which represents a probe mounted on a support strut within a finite sized wind tunnel. Collection efficiencies were generated, using the LEWICE3D code, for four spherical particle sizes, 100, 50, 20, and 5 micron in diameter. It was observed that a reduction in velocity of about 20% occurred, for all cases, as the flow entered the shroud of the probe. The reduction in velocity within the shroud is not indicative of any error in the probe measurement accuracy. Heat transfer results are presented which agree quite well with a correlation for the circular cross section heated elements. Collection efficiency results indicate a reduction in collection efficiency as particle size is reduced. The reduction with particle size is expected, however, the results tended to be lower than the previous results generated for isolated two-dimensional elements. The deviation from the two-dimensional results is more pronounced for the smaller particles and is likely due to the reduced flow within the protective shroud. As particle size increases differences between the two-dimensional and three dimensional results become negligible. Taken as a group, the total collection efficiency of the elements including the effects of the shroud has been shown to be in the range of 0.93 to 0.99 for particles above 20 microns. The 3D model has improved the estimated collection efficiency for smaller particles where errors in previous estimates were more significant.

Water Content↗
NASA NTRS2012

Photosystem Trap Energies and Spectrally-Dependent Energy-Storage Efficiencies in the Chl d-Utilizing Cyanobacterium, Acaryochloris Marina

Acaryochloris marina is the only species known to utilize chlorophyll (Chl) d as a principal photopigment. The peak absorption wavelength of Chl d is redshifted approx. 40 nm in vivo relative to Chl a, enabling this cyanobacterium to perform oxygenic phototrophy in niche environments enhanced in far-red light. We present measurements of the in vivo energy-storage (E-S) efficiency of photosynthesis in A. marina, obtained using pulsed photoacoustics (PA) over a 90-nm range of excitation wavelengths in the red and far-red. Together with modeling results, these measurements provide the first direct observation of the trap energies of PSI and PSII, and also the photosystem-specific contributions to the total E-S efficiency. We find the maximum observed efficiency in A. marina (40+/-1% at 735 nm) is higher than in the Chl a cyanobacterium Synechococcus leopoliensis (35+/-1% at 690 nm). The efficiency at peak absorption wavelength is also higher in A. marina (36+/-1% at 710 nm vs. 31+/-1% at 670 nm). In both species, the trap efficiencies are approx. 40% (PSI) and approx. 30% (PSII). The PSI trap in A. marina is found to lie at 740+/-5 nm, in agreement with the value inferred from spectroscopic methods. The best fit of the model to the PA data identifies the PSII trap at 723+/-3 nm, supporting the view that the primary electron-donor is Chl d, probably at the accessory (ChlD1) site. A decrease in efficiency beyond the trap wavelength, consistent with uphill energy transfer, is clearly observed and fit by the model. These results demonstrate that the E-S efficiency in A. marina is not thermodynamically limited, suggesting that oxygenic photosynthesis is viable in even redder light environments.

Mielke, Steven P.↗
NASA NTRS2017

Luminous Efficiency of Hypervelocity Meteoroid Impacts on the Moon Derived from the 2015 Geminid Meteor Shower

Meteoroids cannot be observed directly because of their small size. In-situ measurements of the meteoroid environment are rare and have very small collecting areas. The Moon, in contrast, has a large collecting area and therefore can be used as a large meteoroid detector for gram-kilogram sized particles. Meteoroids striking the Moon create an impact flash observable by Earth-based telescopes. Their kinetic energy is converted to luminous energy with some unknown luminous efficiency η(v), which is likely a function of meteoroid velocity (among other factors). This luminous efficiency is imperative to calculating the kinetic energy and mass of the meteoroid, as well as meteoroid fluxes, and it cannot be determined in the laboratory at meteoroid speeds and sizes due to mechanical constraints. Since laboratory simulations fail to resolve the luminous efficiency problem, observations of the impact flash itself must be utilized. Meteoroids associated with specific meteor showers have known speed and direction, which simplifies the determination of the luminous efficiency. NASA has routinely monitored the Moon for impact flashes since early 2006 [1]. During this time, several meteor showers have produced multiple impact flashes on the Moon, yielding a sufficient sample of impact flashes with which to perform a luminous efficiency analysis similar to that outlined in Bellot Rubio et al. [2, 3] and further described by Moser et al. [4], utilizing Earth-based measurements of the shower flux and mass index. The Geminid meteor shower has produced the most impact flashes in the NASA dataset to date with over 80 detections. More than half of these Geminids were recorded in 2015 (locations pictured in Fig. 1), and may represent the largest single-shower impact flash sample known. This work analyzes the 2015 Geminid lunar impacts and calculates their luminous efficiency. The luminous efficiency is then applied to calculate the kinetic energies and mass-es of these shower meteoroids.

Moser, D. E.↗
NASA NTRSDate not supplied

Efficient Mirrorlet Array Based Integral Field Spectrometer (IFS) for HWO

Habitable World Observatory (HWO) is a NASA strategic mission recommended by 2020 astronomical decadal survey. Integral field spectrometers play an important role to verify if the observed exoplanet is habitable. The traditional lenslet array based IFS has the advantage of simplicity and compactness. However, it does not use detector pixels efficiently in order to prevent wavelength crosstalk among adjacent spectra. The efficient lenslet/mirrorlet IFS combines the advantages from both lenslet based and imager slicer based IFSes—keeping lenslet IFS’s simplicity and compactness, and adding slicer IFS’s detector efficiency. This paper discusses the principle of efficient lenslet/mirrorlet IFS, design philosophy, and efficient spectral trace layout ideals. It uses HWO NIR IFS requirement as an example to provide an efficiency mirrorlet IFS optical design. The high detector efficiency not only reduces needed detector pixel numbers, but also reduces the high communication rate demanding for such a large multiple instrument mission.

Qian Gong↗
DOE OSTI2022

How to Build a Connected Community: Policies to Promote Grid-interactive Efficient Buildings and Demand Flexibility

Efficient, connected, grid-interactive, smart, and flexible buildings are key to decarbonizing the U.S. energy economy, optimizing energy use, reducing electric consumers’ bills, integrating variable renewable energy resources, and improving the reliability and performance of the nation’s electricity grids. Such grid-interactive efficient buildings have high levels of energy efficiency layered with other distributed energy resources (DERs) and intelligent controls to provide demand flexibility. Policy support is unfolding at the federal, state, and local levels to transform homes and workplaces into state-of-the-art energy-efficient buildings and community-level grid services. This paper starts by describing the potential benefits. Next, it highlights existing policies — with a focus on state-level actions — that support grid-interactive efficient building deployment and demand flexibility. Finally, it identifies current trends and gaps, policies and programs that promote grid-interactive efficient buildings, and aggregations of grid-interactive efficient buildings referred to as Virtual Power Plants.

Schwartz, Lisa C↗
DOE OSTI2026

HARMONY: Large-Scale Architecture Search for Efficient Hybrid Language Models

As large language models scale to trillions of parameters, their computational and memory requirements present critical challenges for efficient training and deployment. While Mixture of Experts (MoE) architectures enable efficient scaling through sparse parameter activation, and state-space models like Mamba offer linear-time complexity, principled methods for combining these paradigms remain undeveloped. We introduce HARMONY (Hybrid Architecture Research for Mamba, Optimized with Neural efficiencY), a multi-objective evolutionary neural architecture search framework for discovering efficient hybrid language models that integrate Transformer attention mechanisms, Mixture-of-Experts routing, and Mamba state-space components. Through large-scale distributed search using 16,384 MI250X GPUs on the Frontier supercomputer, HARMONY explores a comprehensive design space encompassing six attention variants (MHA, MQA, GQA, MLA, SWA, and Mamba-2), variable MoE configurations with both routed and shared experts, and extensive Mamba hyperparameters. Our framework discovers heterogeneous architectures that balance training performance with computational efficiency through multi-objective optimization incorporating latency penalties and fitness-based selection. Analysis of discovered architectures reveals that optimal hybrid designs favor heterogeneous component mixing rather than homogeneous patterns, with Mamba-2 and Multi-Head Latent Attention (MLA) emerging as preferred mechanisms. Discovered architectures demonstrate superior training efficiency: our best configuration achieves a final perplexity of 1.0874 with 2.38B parameters while processing 4,320 tokens/second, outperforming significantly larger manually designed models. Full-scale evaluation shows HARMONY's top architectures achieve better loss trajectories than equivalently-sized models using state-of-the-art configurations including Mixtral, Jamba, and Samba. Additionally, we demonstrate 91% weak scaling efficiency when training discovered 36B-parameter models across 1,024 GPUs. HARMONY is released as an open framework with comprehensive tools for building and training hybrid models using expert-data-pipeline parallelism, democratizing access to automated architecture design for next-generation language models.

Herron, Emily [ORNL] (ORCID:0000000273008172)↗